mirror/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-01-01 12:08:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

Better accuracy for FREST/FRSQEST (#13863)

Browse Source
This commit is contained in:
RipleyTom 2023-05-15 18:20:47 +02:00 committed by GitHub
parent 514ef9a9c5
commit f11770b88b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 12 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
rpcs3/Emu/Cell/SPURecompiler.cpp
View File

@ -8865,11 +8865,10 @@ public:
register_intrinsic("spu_frest", [&](llvm::CallInst* ci)
{
const auto a = value<f32[4]>(ci->getOperand(0));
const auto acc_result = fsplat<f32[4]>(1.0) / a;
// Determines accuracy penalty, frest result is always slightly closer to 0 than actual value and provides ~12 bits accuracy
const auto acc_penalty = fsplat<f32[4]>(0x1p-13f) * acc_result;
// Gives accuracy penalty, frest result is within one newton-raphson iteration for accuracy
const auto approx_result = fsplat<f32[4]>(0.999875069f) / a;
// Zeroes the last 11 bytes of the mantissa so FI calculations end up correct if needed
return bitcast<f32[4]>(bitcast<u32[4]>(acc_result - acc_penalty) & splat<u32[4]>(0xFFFFF800));
return bitcast<f32[4]>(bitcast<u32[4]>(approx_result) & splat<u32[4]>(0xFFFFF800));
});
}
else
@ -8905,11 +8904,10 @@ public:
register_intrinsic("spu_frsqest", [&](llvm::CallInst* ci)
{
const auto a = value<f32[4]>(ci->getOperand(0));
const auto acc_result = fsplat<f32[4]>(1.0) / fsqrt(fabs(a));
// Determines accuracy penalty, frsqest result is always slightly closer to 0 than actual value and provides ~12 bits accuracy
const auto acc_penalty = fsplat<f32[4]>(0x1p-13f) * acc_result;
// Gives accuracy penalty, frsqest result is within one newton-raphson iteration for accuracy
const auto approx_result = fsplat<f32[4]>(0.999763668f) / fsqrt(fabs(a));
// Zeroes the last 11 bytes of the mantissa so FI calculations end up correct if needed
return bitcast<f32[4]>(bitcast<u32[4]>(acc_result - acc_penalty) & splat<u32[4]>(0xFFFFF800));
return bitcast<f32[4]>(bitcast<u32[4]>(approx_result) & splat<u32[4]>(0xFFFFF800));
});
}
else
@ -9648,19 +9646,17 @@ public:
register_intrinsic("spu_re", [&](llvm::CallInst* ci)
{
const auto a = value<f32[4]>(ci->getOperand(0));
const auto acc_result = fsplat<f32[4]>(1.0) / a;
// Determines accuracy penalty, frest result is always slightly closer to 0 than actual value and provides ~12 bits accuracy
const auto acc_penalty = fsplat<f32[4]>(0x1p-13f) * acc_result;
return acc_result - acc_penalty;
// Gives accuracy penalty, frest result is within one newton-raphson iteration for accuracy
const auto approx_result = fsplat<f32[4]>(0.999875069f) / a;
return approx_result;
});
register_intrinsic("spu_rsqrte", [&](llvm::CallInst* ci)
{
const auto a = value<f32[4]>(ci->getOperand(0));
const auto acc_result = fsplat<f32[4]>(1.0) / fsqrt(fabs(a));
// Determines accuracy penalty, frsqest result is always slightly closer to 0 than actual value and provides ~12 bits accuracy
const auto acc_penalty = fsplat<f32[4]>(0x1p-13f) * acc_result;
return acc_result - acc_penalty;
// Gives accuracy penalty, frsqest result is within one newton-raphson iteration for accuracy
const auto approx_result = fsplat<f32[4]>(0.999763668f) / fsqrt(fabs(a));
return approx_result;
});
}
else